4 * Builtin report command: Analyze the perf.data input file,
5 * look up and read DSOs and symbol information and display
6 * a histogram of results, along various sorting keys.
10 #include "util/util.h"
12 #include "util/annotate.h"
13 #include "util/color.h"
14 #include <linux/list.h>
15 #include "util/cache.h"
16 #include <linux/rbtree.h>
17 #include "util/symbol.h"
18 #include "util/callchain.h"
19 #include "util/strlist.h"
20 #include "util/values.h"
23 #include "util/debug.h"
24 #include "util/evlist.h"
25 #include "util/evsel.h"
26 #include "util/header.h"
27 #include "util/session.h"
28 #include "util/tool.h"
30 #include "util/parse-options.h"
31 #include "util/parse-events.h"
33 #include "util/thread.h"
34 #include "util/sort.h"
35 #include "util/hist.h"
36 #include "arch/common.h"
38 #include <linux/bitmap.h>
41 struct perf_tool tool;
42 struct perf_session *session;
43 bool force, use_tui, use_gtk, use_stdio;
45 bool dont_use_callchains;
48 bool inverted_callchain;
49 struct perf_read_values show_threads_values;
50 const char *pretty_printing_style;
51 symbol_filter_t annotate_init;
53 const char *symbol_filter_str;
54 DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
57 static int perf_report__add_branch_hist_entry(struct perf_tool *tool,
58 struct addr_location *al,
59 struct perf_sample *sample,
60 struct perf_evsel *evsel,
61 struct machine *machine)
63 struct perf_report *rep = container_of(tool, struct perf_report, tool);
64 struct symbol *parent = NULL;
67 struct hist_entry *he;
68 struct branch_info *bi, *bx;
70 if ((sort__has_parent || symbol_conf.use_callchain)
71 && sample->callchain) {
72 err = machine__resolve_callchain(machine, evsel, al->thread,
78 bi = machine__resolve_bstack(machine, al->thread,
79 sample->branch_stack);
83 for (i = 0; i < sample->branch_stack->nr; i++) {
84 if (rep->hide_unresolved && !(bi[i].from.sym && bi[i].to.sym))
87 * The report shows the percentage of total branches captured
88 * and not events sampled. Thus we use a pseudo period of 1.
90 he = __hists__add_branch_entry(&evsel->hists, al, parent,
93 struct annotation *notes;
96 if (bx->from.sym && use_browser == 1 && sort__has_sym) {
97 notes = symbol__annotation(bx->from.sym);
99 && symbol__alloc_hist(bx->from.sym) < 0)
102 err = symbol__inc_addr_samples(bx->from.sym,
110 if (bx->to.sym && use_browser == 1 && sort__has_sym) {
111 notes = symbol__annotation(bx->to.sym);
113 && symbol__alloc_hist(bx->to.sym) < 0)
116 err = symbol__inc_addr_samples(bx->to.sym,
123 evsel->hists.stats.total_period += 1;
124 hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
133 static int perf_evsel__add_hist_entry(struct perf_evsel *evsel,
134 struct addr_location *al,
135 struct perf_sample *sample,
136 struct machine *machine)
138 struct symbol *parent = NULL;
140 struct hist_entry *he;
142 if ((sort__has_parent || symbol_conf.use_callchain) && sample->callchain) {
143 err = machine__resolve_callchain(machine, evsel, al->thread,
149 he = __hists__add_entry(&evsel->hists, al, parent, sample->period);
153 if (symbol_conf.use_callchain) {
154 err = callchain_append(he->callchain,
161 * Only in the newt browser we are doing integrated annotation,
162 * so we don't allocated the extra space needed because the stdio
163 * code will not use it.
165 if (he->ms.sym != NULL && use_browser == 1 && sort__has_sym) {
166 struct annotation *notes = symbol__annotation(he->ms.sym);
168 assert(evsel != NULL);
171 if (notes->src == NULL && symbol__alloc_hist(he->ms.sym) < 0)
174 err = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
177 evsel->hists.stats.total_period += sample->period;
178 hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
184 static int process_sample_event(struct perf_tool *tool,
185 union perf_event *event,
186 struct perf_sample *sample,
187 struct perf_evsel *evsel,
188 struct machine *machine)
190 struct perf_report *rep = container_of(tool, struct perf_report, tool);
191 struct addr_location al;
193 if (perf_event__preprocess_sample(event, machine, &al, sample,
194 rep->annotate_init) < 0) {
195 fprintf(stderr, "problem processing %d event, skipping it.\n",
200 if (al.filtered || (rep->hide_unresolved && al.sym == NULL))
203 if (rep->cpu_list && !test_bit(sample->cpu, rep->cpu_bitmap))
206 if (sort__branch_mode == 1) {
207 if (perf_report__add_branch_hist_entry(tool, &al, sample,
209 pr_debug("problem adding lbr entry, skipping event\n");
214 al.map->dso->hit = 1;
216 if (perf_evsel__add_hist_entry(evsel, &al, sample, machine)) {
217 pr_debug("problem incrementing symbol period, skipping event\n");
224 static int process_read_event(struct perf_tool *tool,
225 union perf_event *event,
226 struct perf_sample *sample __maybe_unused,
227 struct perf_evsel *evsel,
228 struct machine *machine __maybe_unused)
230 struct perf_report *rep = container_of(tool, struct perf_report, tool);
232 if (rep->show_threads) {
233 const char *name = evsel ? perf_evsel__name(evsel) : "unknown";
234 perf_read_values_add_value(&rep->show_threads_values,
235 event->read.pid, event->read.tid,
241 dump_printf(": %d %d %s %" PRIu64 "\n", event->read.pid, event->read.tid,
242 evsel ? perf_evsel__name(evsel) : "FAIL",
248 /* For pipe mode, sample_type is not currently set */
249 static int perf_report__setup_sample_type(struct perf_report *rep)
251 struct perf_session *self = rep->session;
252 u64 sample_type = perf_evlist__sample_type(self->evlist);
254 if (!self->fd_pipe && !(sample_type & PERF_SAMPLE_CALLCHAIN)) {
255 if (sort__has_parent) {
256 ui__error("Selected --sort parent, but no "
257 "callchain data. Did you call "
258 "'perf record' without -g?\n");
261 if (symbol_conf.use_callchain) {
262 ui__error("Selected -g but no callchain data. Did "
263 "you call 'perf record' without -g?\n");
266 } else if (!rep->dont_use_callchains &&
267 callchain_param.mode != CHAIN_NONE &&
268 !symbol_conf.use_callchain) {
269 symbol_conf.use_callchain = true;
270 if (callchain_register_param(&callchain_param) < 0) {
271 ui__error("Can't register callchain params.\n");
276 if (sort__branch_mode == 1) {
277 if (!self->fd_pipe &&
278 !(sample_type & PERF_SAMPLE_BRANCH_STACK)) {
279 ui__error("Selected -b but no branch data. "
280 "Did you call perf record without -b?\n");
288 extern volatile int session_done;
290 static void sig_handler(int sig __maybe_unused)
295 static size_t hists__fprintf_nr_sample_events(struct hists *self,
296 const char *evname, FILE *fp)
300 unsigned long nr_samples = self->stats.nr_events[PERF_RECORD_SAMPLE];
301 u64 nr_events = self->stats.total_period;
302 struct perf_evsel *evsel = hists_to_evsel(self);
304 size_t size = sizeof(buf);
306 if (symbol_conf.event_group && evsel->nr_members > 1) {
307 struct perf_evsel *pos;
309 perf_evsel__group_desc(evsel, buf, size);
312 for_each_group_member(pos, evsel) {
313 nr_samples += pos->hists.stats.nr_events[PERF_RECORD_SAMPLE];
314 nr_events += pos->hists.stats.total_period;
318 nr_samples = convert_unit(nr_samples, &unit);
319 ret = fprintf(fp, "# Samples: %lu%c", nr_samples, unit);
321 ret += fprintf(fp, " of event '%s'", evname);
323 ret += fprintf(fp, "\n# Event count (approx.): %" PRIu64, nr_events);
324 return ret + fprintf(fp, "\n#\n");
327 static int perf_evlist__tty_browse_hists(struct perf_evlist *evlist,
328 struct perf_report *rep,
331 struct perf_evsel *pos;
333 list_for_each_entry(pos, &evlist->entries, node) {
334 struct hists *hists = &pos->hists;
335 const char *evname = perf_evsel__name(pos);
337 if (symbol_conf.event_group &&
338 !perf_evsel__is_group_leader(pos))
341 hists__fprintf_nr_sample_events(hists, evname, stdout);
342 hists__fprintf(hists, true, 0, 0, stdout);
343 fprintf(stdout, "\n\n");
346 if (sort_order == default_sort_order &&
347 parent_pattern == default_parent_pattern) {
348 fprintf(stdout, "#\n# (%s)\n#\n", help);
350 if (rep->show_threads) {
351 bool style = !strcmp(rep->pretty_printing_style, "raw");
352 perf_read_values_display(stdout, &rep->show_threads_values,
354 perf_read_values_destroy(&rep->show_threads_values);
361 static int __cmd_report(struct perf_report *rep)
365 struct perf_session *session = rep->session;
366 struct perf_evsel *pos;
367 struct map *kernel_map;
368 struct kmap *kernel_kmap;
369 const char *help = "For a higher level overview, try: perf report --sort comm,dso";
371 signal(SIGINT, sig_handler);
374 ret = perf_session__cpu_bitmap(session, rep->cpu_list,
380 if (use_browser <= 0)
381 perf_session__fprintf_info(session, stdout, rep->show_full_info);
383 if (rep->show_threads)
384 perf_read_values_init(&rep->show_threads_values);
386 ret = perf_report__setup_sample_type(rep);
390 ret = perf_session__process_events(session, &rep->tool);
394 kernel_map = session->machines.host.vmlinux_maps[MAP__FUNCTION];
395 kernel_kmap = map__kmap(kernel_map);
396 if (kernel_map == NULL ||
397 (kernel_map->dso->hit &&
398 (kernel_kmap->ref_reloc_sym == NULL ||
399 kernel_kmap->ref_reloc_sym->addr == 0))) {
401 "As no suitable kallsyms nor vmlinux was found, kernel samples\n"
402 "can't be resolved.";
405 const struct dso *kdso = kernel_map->dso;
406 if (!RB_EMPTY_ROOT(&kdso->symbols[MAP__FUNCTION])) {
407 desc = "If some relocation was applied (e.g. "
408 "kexec) symbols may be misresolved.";
413 "Kernel address maps (/proc/{kallsyms,modules}) were restricted.\n\n"
414 "Check /proc/sys/kernel/kptr_restrict before running 'perf record'.\n\n%s\n\n"
415 "Samples in kernel modules can't be resolved as well.\n\n",
420 perf_session__fprintf(session, stdout);
423 perf_session__fprintf_dsos(session, stdout);
426 perf_session__fprintf_nr_events(session, stdout);
431 list_for_each_entry(pos, &session->evlist->entries, node) {
432 struct hists *hists = &pos->hists;
435 hists->symbol_filter_str = rep->symbol_filter_str;
437 hists__collapse_resort(hists);
438 nr_samples += hists->stats.nr_events[PERF_RECORD_SAMPLE];
440 /* Non-group events are considered as leader */
441 if (symbol_conf.event_group &&
442 !perf_evsel__is_group_leader(pos)) {
443 struct hists *leader_hists = &pos->leader->hists;
445 hists__match(leader_hists, hists);
446 hists__link(leader_hists, hists);
450 if (nr_samples == 0) {
451 ui__error("The %s file has no samples!\n", session->filename);
455 list_for_each_entry(pos, &session->evlist->entries, node)
456 hists__output_resort(&pos->hists);
458 if (use_browser > 0) {
459 if (use_browser == 1) {
460 perf_evlist__tui_browse_hists(session->evlist, help,
462 &session->header.env);
463 } else if (use_browser == 2) {
464 perf_evlist__gtk_browse_hists(session->evlist, help,
468 perf_evlist__tty_browse_hists(session->evlist, rep, help);
472 * Speed up the exit process, for large files this can
473 * take quite a while.
475 * XXX Enable this when using valgrind or if we ever
476 * librarize this command.
478 * Also experiment with obstacks to see how much speed
481 * perf_session__delete(session);
487 parse_callchain_opt(const struct option *opt, const char *arg, int unset)
489 struct perf_report *rep = (struct perf_report *)opt->value;
497 rep->dont_use_callchains = true;
501 symbol_conf.use_callchain = true;
506 tok = strtok((char *)arg, ",");
510 /* get the output mode */
511 if (!strncmp(tok, "graph", strlen(arg)))
512 callchain_param.mode = CHAIN_GRAPH_ABS;
514 else if (!strncmp(tok, "flat", strlen(arg)))
515 callchain_param.mode = CHAIN_FLAT;
517 else if (!strncmp(tok, "fractal", strlen(arg)))
518 callchain_param.mode = CHAIN_GRAPH_REL;
520 else if (!strncmp(tok, "none", strlen(arg))) {
521 callchain_param.mode = CHAIN_NONE;
522 symbol_conf.use_callchain = false;
530 /* get the min percentage */
531 tok = strtok(NULL, ",");
535 callchain_param.min_percent = strtod(tok, &endptr);
539 /* get the print limit */
540 tok2 = strtok(NULL, ",");
544 if (tok2[0] != 'c') {
545 callchain_param.print_limit = strtoul(tok2, &endptr, 0);
546 tok2 = strtok(NULL, ",");
551 /* get the call chain order */
552 if (!strcmp(tok2, "caller"))
553 callchain_param.order = ORDER_CALLER;
554 else if (!strcmp(tok2, "callee"))
555 callchain_param.order = ORDER_CALLEE;
559 if (callchain_register_param(&callchain_param) < 0) {
560 fprintf(stderr, "Can't register callchain params\n");
567 parse_branch_mode(const struct option *opt __maybe_unused,
568 const char *str __maybe_unused, int unset)
570 sort__branch_mode = !unset;
574 int cmd_report(int argc, const char **argv, const char *prefix __maybe_unused)
576 struct perf_session *session;
578 bool has_br_stack = false;
580 char callchain_default_opt[] = "fractal,0.5,callee";
581 const char * const report_usage[] = {
582 "perf report [<options>]",
585 struct perf_report report = {
587 .sample = process_sample_event,
588 .mmap = perf_event__process_mmap,
589 .comm = perf_event__process_comm,
590 .exit = perf_event__process_exit,
591 .fork = perf_event__process_fork,
592 .lost = perf_event__process_lost,
593 .read = process_read_event,
594 .attr = perf_event__process_attr,
595 .event_type = perf_event__process_event_type,
596 .tracing_data = perf_event__process_tracing_data,
597 .build_id = perf_event__process_build_id,
598 .ordered_samples = true,
599 .ordering_requires_timestamps = true,
601 .pretty_printing_style = "normal",
603 const struct option options[] = {
604 OPT_STRING('i', "input", &input_name, "file",
606 OPT_INCR('v', "verbose", &verbose,
607 "be more verbose (show symbol address, etc)"),
608 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
609 "dump raw trace in ASCII"),
610 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
611 "file", "vmlinux pathname"),
612 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
613 "file", "kallsyms pathname"),
614 OPT_BOOLEAN('f', "force", &report.force, "don't complain, do it"),
615 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
616 "load module symbols - WARNING: use only with -k and LIVE kernel"),
617 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
618 "Show a column with the number of samples"),
619 OPT_BOOLEAN('T', "threads", &report.show_threads,
620 "Show per-thread event counters"),
621 OPT_STRING(0, "pretty", &report.pretty_printing_style, "key",
622 "pretty printing style key: normal raw"),
623 OPT_BOOLEAN(0, "tui", &report.use_tui, "Use the TUI interface"),
624 OPT_BOOLEAN(0, "gtk", &report.use_gtk, "Use the GTK2 interface"),
625 OPT_BOOLEAN(0, "stdio", &report.use_stdio,
626 "Use the stdio interface"),
627 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
628 "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline,"
629 " dso_to, dso_from, symbol_to, symbol_from, mispredict"),
630 OPT_BOOLEAN(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
631 "Show sample percentage for different cpu modes"),
632 OPT_STRING('p', "parent", &parent_pattern, "regex",
633 "regex filter to identify parent, see: '--sort parent'"),
634 OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
635 "Only display entries with parent-match"),
636 OPT_CALLBACK_DEFAULT('g', "call-graph", &report, "output_type,min_percent[,print_limit],call_order",
637 "Display callchains using output_type (graph, flat, fractal, or none) , min percent threshold, optional print limit and callchain order. "
638 "Default: fractal,0.5,callee", &parse_callchain_opt, callchain_default_opt),
639 OPT_BOOLEAN('G', "inverted", &report.inverted_callchain,
640 "alias for inverted call graph"),
641 OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
642 "only consider symbols in these dsos"),
643 OPT_STRING('c', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
644 "only consider symbols in these comms"),
645 OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
646 "only consider these symbols"),
647 OPT_STRING(0, "symbol-filter", &report.symbol_filter_str, "filter",
648 "only show symbols that (partially) match with this filter"),
649 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
651 "don't try to adjust column width, use these fixed values"),
652 OPT_STRING('t', "field-separator", &symbol_conf.field_sep, "separator",
653 "separator for columns, no spaces will be added between "
654 "columns '.' is reserved."),
655 OPT_BOOLEAN('U', "hide-unresolved", &report.hide_unresolved,
656 "Only display entries resolved to a symbol"),
657 OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
658 "Look for files with symbols relative to this directory"),
659 OPT_STRING('C', "cpu", &report.cpu_list, "cpu",
660 "list of cpus to profile"),
661 OPT_BOOLEAN('I', "show-info", &report.show_full_info,
662 "Display extended information about perf.data file"),
663 OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
664 "Interleave source code with assembly code (default)"),
665 OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
666 "Display raw encoding of assembly instructions (default)"),
667 OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
668 "Specify disassembler style (e.g. -M intel for intel syntax)"),
669 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
670 "Show a column with the sum of periods"),
671 OPT_CALLBACK_NOOPT('b', "branch-stack", &sort__branch_mode, "",
672 "use branch records for histogram filling", parse_branch_mode),
673 OPT_STRING(0, "objdump", &objdump_path, "path",
674 "objdump binary to use for disassembly and annotations"),
678 argc = parse_options(argc, argv, options, report_usage, 0);
680 if (report.use_stdio)
682 else if (report.use_tui)
684 else if (report.use_gtk)
687 if (report.inverted_callchain)
688 callchain_param.order = ORDER_CALLER;
690 if (!input_name || !strlen(input_name)) {
691 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
694 input_name = "perf.data";
696 session = perf_session__new(input_name, O_RDONLY,
697 report.force, false, &report.tool);
701 report.session = session;
703 has_br_stack = perf_header__has_feat(&session->header,
704 HEADER_BRANCH_STACK);
706 if (sort__branch_mode == -1 && has_br_stack)
707 sort__branch_mode = 1;
709 /* sort__branch_mode could be 0 if --no-branch-stack */
710 if (sort__branch_mode == 1) {
712 * if no sort_order is provided, then specify
713 * branch-mode specific order
715 if (sort_order == default_sort_order)
716 sort_order = "comm,dso_from,symbol_from,"
721 if (strcmp(input_name, "-") != 0)
725 perf_hpp__column_enable(PERF_HPP__OVERHEAD);
729 setup_sorting(report_usage, options);
732 * Only in the newt browser we are doing integrated annotation,
733 * so don't allocate extra space that won't be used in the stdio
736 if (use_browser == 1 && sort__has_sym) {
737 symbol_conf.priv_size = sizeof(struct annotation);
738 report.annotate_init = symbol__annotate_init;
740 * For searching by name on the "Browse map details".
741 * providing it only in verbose mode not to bloat too
742 * much struct symbol.
746 * XXX: Need to provide a less kludgy way to ask for
747 * more space per symbol, the u32 is for the index on
749 * See symbol__browser_index.
751 symbol_conf.priv_size += sizeof(u32);
752 symbol_conf.sort_by_name = true;
756 if (symbol__init() < 0)
759 if (parent_pattern != default_parent_pattern) {
760 if (sort_dimension__add("parent") < 0)
764 * Only show the parent fields if we explicitly
765 * sort that way. If we only use parent machinery
766 * for filtering, we don't want it.
768 if (!strstr(sort_order, "parent"))
769 sort_parent.elide = 1;
771 symbol_conf.exclude_other = false;
775 * Special case: if there's an argument left then assume that
776 * it's a symbol filter:
779 usage_with_options(report_usage, options);
781 report.symbol_filter_str = argv[0];
784 sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
786 if (sort__branch_mode == 1) {
787 sort_entry__setup_elide(&sort_dso_from, symbol_conf.dso_from_list, "dso_from", stdout);
788 sort_entry__setup_elide(&sort_dso_to, symbol_conf.dso_to_list, "dso_to", stdout);
789 sort_entry__setup_elide(&sort_sym_from, symbol_conf.sym_from_list, "sym_from", stdout);
790 sort_entry__setup_elide(&sort_sym_to, symbol_conf.sym_to_list, "sym_to", stdout);
792 sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
793 sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
796 ret = __cmd_report(&report);
798 perf_session__delete(session);